reindex resource selection update slows down as reindex progresses (postgres)

[punktilious]

Describe the bug
The reindex operation uses a statement like the following to pick the next resource to reindex. This is intended to avoid any concurrency issues. Unfortunately, although the index on reindex_tstamp is DESC, the scan ends up being a forward scan and must hit more and more blocks in order to find the first resource matching the filter predicate:

UPDATE fhirdata.logical_resources    
   SET reindex_tstamp = '2020-12-10T00:00:00Z'     
WHERE logical_resource_id = (
	SELECT lr.logical_resource_id          
  	  FROM fhirdata.logical_resources lr        
	 WHERE lr.reindex_tstamp < '2020-12-10T00:00:00Z'
  ORDER BY lr.reindex_tstamp DESC    
       FOR UPDATE SKIP LOCKED LIMIT 1) 
RETURNING logical_resource_id, resource_type_id, logical_id, reindex_txid;

Removing the DESC on the order by reduces the number of shared buffers hit from 50,000 to 7, and reduces the execution time from 900ms to 0.018ms according to EXPLAIN (ANALYZE,BUFFERS):

UPDATE fhirdata.logical_resources    
   SET reindex_tstamp = '2020-12-10T00:00:00Z'     
WHERE logical_resource_id = (
	SELECT lr.logical_resource_id          
  	  FROM fhirdata.logical_resources lr        
	 WHERE lr.reindex_tstamp < '2020-12-10T00:00:00Z'
  ORDER BY lr.reindex_tstamp   
       FOR UPDATE SKIP LOCKED LIMIT 1) 
RETURNING logical_resource_id, resource_type_id, logical_id, reindex_txid;

Explain output for the SELECT part of the statement:

"Limit  (cost=0.57..1.65 rows=1 width=16) (actual time=0.038..0.045 rows=1 loops=1)"
"  Buffers: shared hit=7"
"  ->  Index Scan Backward using idx_logical_resources_rits on logical_resources lr  (cost=0.57..113322070.36 rows=104493008 width=16) (actual time=0.034..0.036 rows=1 loops=1)"
"        Index Cond: (reindex_tstamp < '2020-12-10 00:00:00'::timestamp without time zone)"
"        Buffers: shared hit=7"
"Planning Time: 0.127 ms"
"Execution Time: 0.081 ms"